Volume 2, Issue 9, September 2017 International Journal of Innovative Science and Research Technology
ISSN No: - 2456 2165
Electricity Load Prediction Using Artificial Neural
Network By Back Propagation
Joe Joseph, Nithin James, Sebastian Sachin, Varkey Vincent, Sminu Izudheen
Department of Computer Science and Engineering, Rajagiri School of Engineering and Technology
Rajagiri Valley, Kakkanad, Kerala, 682039, India
Abstract This paper presents a data mining approach to
forecast the power consumption of a geographical region
based on the meteorological data. Artificial Neural
Network using Back Propagation is the data mining
approach adopted here. The research is conducted using
meteorological and load consumption data in Kerala
region over the period of 2011- 2012. The historical data
containing temperature, humidity, public holidays and the
daily load consumption values are used for training the
system. A predictive model is then built which gather
information from the training data set and generates
knowledge for predicting the demand of the region. Results
on statistical significance tests assert that the proposed
method can be used as an effective model to study the load
consumption pattern which in turn helps to forecast the
electricity demand of a region.
Keywords Data Mining;Artificial Neural Network;Back
Propagation.
I. INTRODUCTION
For best utilization of electricity, power prediction plays a
significant role in energy management. Load forecasting helps
a country to plan, schedule and manage its energy generation
to satisfy the growing needs. Forecasting demand helps to plan
the means to satisfy the demand, efficient use of existing
energy sources, new infrastructure to be build. It also helps in
meeting the fluctuating demand, giving information about
consumption pattern.
Load forecasting is challenging because it may depend upon a
number of factors such as the day, season, meteorological data
and holidays. Some researches may find all data to be useful
but for some it may not be. Forecasting may depend on the
attributes that are considered. Load forecasting can be
classified into short term, medium term, or long term. There
are a number of methods for electricity forecast which
includes fuzzy logic, Artificial Neural Network (ANN),
Support Vector Machine (SVM), Regression etc.
We adopt ANN with back propagation for load prediction. In
recent years, ANN as the consequence of their predominant
capability in defining the nonlinear topography between
different nonlinear variables has become an approved
procedure in the load forecasting area. We are designing a
system for short term forecasting. The input considered are
IJISRT17SP221 www.ijisrt.com
meteorological and holiday values. The meteorological value
includes minimum temperature and humidity, maximum
temperature and humidity. The system is built with two
models. First one is the training model which trains the system
and the latter is the prediction model which forecast the
required load.
In the training phase the preprocessed data is fed to the
system. Preprocessing the data is an important step in data
mining as the input data greatly affects the performance of the
system. Outlier removal and the min-max normalization are
the two preprocessing techniques adopted. The training
percentage can be selected by the user. The system will be
trained better if more data is presented to it. The accuracy of
the system is then checked with the remaining data known as
the test data.
After the training phase the system is then ready to predict the
load, given the required inputs. A python script is used to get
the meteorological value from the wunderground database for
the next 10 days. User can select the weekday for which the
demand prediction has to be made. Section 4 gives a brief
about the methodology adopted.
II. RELATED WORK
There are a number of researches going on this field. Different
data mining techniques have been adopted to predict the load.
Omer F. Demirel, Selim Zaim has used both SVM and neural
network to forecast electricity demand for Turkey. Hoda K.
Mohamed, Soliman M. El-Debeiky, Hassan M.Mahmoud,
Khaled M. El Destawy presented a model for forecasting long
term electricity load in Egyptian Electrical Network [6]. The
demand is attributed to a total of eleven parameters-total
energy consumption, total energy generation, energy losses,
load factor, GDP, electricity price, population, temperature,
humidity.
Tso[5]presents three modeling techniques for the prediction of
electricity energy consumption in two year seasons: summer
and winter. The three predictive modeling techniques are
multiple regressions, a neural network and decision tree
models. The base ARIMA model with just one seasonal
pattern can be extended for the case of multiple seasons. An
example of such an extension was presented in [3]. A
combinatorial problem of selecting appropriate model orders is
298
Volume 2, Issue 9, September 2017
an inconvenience in the time series modeling using multiple
seasonal ARIMA. Another disadvantage is the linear character
of the ARIMA model.
A neural network with back propagation momentum training
algorithm was proposed in the paper Artificial Neural Network
(ANN) trained by the Artificial Immune System (AIS)
presented by M.B. Abdul Hamid and T.K. Abdul Rahman for
load forecasting in order to reduce training time and to
improve convergence speed for short term load forecasting
model [7]. This algorithm has specific benefits such as
accuracy, speed of convergence, economic and historical data
requirement for training etc. The major benefit of this
algorithm over back propagation algorithm is in terms of
improvement in mean average percentage error (MAPE).
P.Sibi, S.Allwyn Jones, P.Siddarth [4] has carried out a
research on different activation functions that can be applied to
the neural network.
III. ANN
Artificial neural networks (ANN) are computational models
inspired by the human brain. At their core, they are comprised
of a large number of connected nodes, each of which performs
a simple mathematical operation. Each node's output is
determined by this operation, as well as a set of parameters
that are specific to that node. By connecting these nodes
together and carefully setting their parameters, very complex
functions can be learned and calculated.
Fig 1 : Neural Network Layout
There are three layers in a neural network. These are input,
hidden and output layer. Only a single input and output layer
is present and any number of hidden layers can be present.
Once an input is presented to the neural network, and a
corresponding desired or target response is set at the output, an
error is composed from the difference of the desired response
and the real system output. The error information is fed back
to the system which makes all adjustments to their parameters
in a systematic fashion. This process is repeated until the
desired output is acceptable.
IJISRT17SP221 www.ijisrt.com
International Journal of Innovative Science and Research Technology
ISSN No: - 2456 2165
IV. METHODOLOGY
A. Data Collection
Two sets of data are collected. One set consists of
meteorological data and the other set is about the states
electricity consumption. The meteorological data is collected
from the database stored by wunderground. This data consists
of minimum and maximum temperature of the day, minimum
and maximum humidity of the day. In addition to it the dates
of holidays are also collected. The electricity consumption
data consists of the daily demand of the state. It is collected
from the Electricity Grid of the state. These data are collected
for the years 2011and 2012.
B. Data Integration
All the data collected i.e. the electricity demand data,
meteorological data and holiday data is combined to form a
single data set. This data is combined based on the
corresponding date.
C. Data Preprocessing
Data pre-processing is an important step in data mining. The
performance of a system is highly dependent on the input data.
So, the data fed to the system should be free of noise and
missing values. In the pre-processing stage we deal with
outlier removal and normalization.
Outlier removal: An outlier is an observation point that is
distant from other observations. An outlier may be due to
variability in the measurement or it may indicate experimental
error.
The interquartile range (IQR) is a measure of variability, based
on dividing a data set into quartiles. Quartiles divide a rank-
ordered data set into four equal parts. The values that separate
parts are called the first, second, and third quartiles; and they
are denoted by Q1, Q2, and Q3, respectively.
Q1 is the "middle" value in the first half of the rank-
ordered data set.
Q2 is the median value in the set.
Q3 is the "middle" value in the second half of the
rank-ordered data set.
IQR=Q3-Q1 (1)
Up=Q3 + (IQR*1.5) (2)
Low=Q1 - (IQR*1.5) (3)
Outliers are those values which are greater than up value and
lesser than Low value. These values are detected and removed.
Min Max normalization: Normalization is a process of
scaling down the attribute values into a specified range. There
are two reasons for normalization. First to eliminate the
influence of one factor over another (i.e. to give features equal
chances), second reason is that the gradient descent which is
299
Volume 2, Issue 9, September 2017 International Journal of Innovative Science and Research Technology
used for back propagation converges faster with normalized
data than with un-normalized data. Here we are using a simple
technique called Min-Max Normalization.
(4)
D. Data Mining
After the preprocessing of the data, now we have data which is
free from missing values, noise and outliers. These data are
now fed to the system for training. The percentage of data for
training can be specified by the user. Usually 80% of data is
given for training. The user interface also provides selection of
preprocessing techniques.
The Fig 2 depicts the system architecture. The system is
provided with 6 inputs. They are minimum temperature and
humidity, maximum temperature and humidity, holiday and
the weekday. The neural network consists of one input layer
Fig 2: System Architecture
The rest of the input i.e. the test set is provided to the
prediction module and tested. Then load is predicted by the
system when providing with the necessary inputs.
E. Visualization of knowledge
Better training occurs when more data are presented to the
system. So for training 80% of data are used. After training,
IJISRT17SP221 www.ijisrt.com
ISSN No: - 2456 2165
with 6 nodes, two hidden layers with 4 and 2 nodes
respectively and an output layer with a single node (Fig 1). As
there is no general rule to find the hidden layers it was fixed
using trial and error method. The activation function used is
log sigmoid function. It is applied to both hidden and output
layer nodes.
Initialization of weights: The weights are initialized randomly
in the range 0-1. Later it gets updated to positive or negative
values as the system progresses.
Termination condition: Here the termination condition is on
reaching the no. of epoch specified and it is within a minimum
error specified. The epoch specified in this work is 100000 and
minimum error is 0.005.
As the preprocessed input is fed the required processing is
done in each node and the weights are updated. The learning
process will terminate based on the termination condition.
Now we have a trained model which is capable of predicting
the load.
the remaining 20% of data is fed to the model for representing
the knowledge. An x-y graph which is plotted using java is
used for this purpose. The graph gives a plot between the
actual load and the load predicted by the model. The x axis
corresponds to the sample number and the y axis corresponds
to the load. The red colored line shows the actual load and the
green colored line shows the predicted load.
300
Volume 2, Issue 9, September 2017 International Journal of Innovative Science and Research Technology
ISSN No: - 2456 2165

[3]. Taylor, J.W.Short-Term Electricity Demand Forecasting

Using Double Seasonal Exponential.
[4]. P.Sibi, S.Allwyn Jones, P.Siddarth, Analysisof Different
Activation Functions Using Back Propagation Neural
Networks, Journal of Theoretical and Applied Information
Technology; Vol. 47 No.3.
[5]. Tso GKF, Yau KKW. Predicting electricity energy
consumption: A comparison of regression analysis,
decision tree and neural networks. Energy 2005; vol 28 pp
1761-1768.
[6]. Hoda K. Mohamed, Soliman M. El-Debeiky, Hassan
M.Mahmoud, Khaled M. El Destawy . Data mining
forelectrical load forecasting in egyptian electrical
network.Computer Engineering
andSystems;IEEE,2006,460-465.
[7]. M. B. Abdul Hamid and T. K. Abdul Rahman. Short Term
Load Forecasting Using an Artificial Neural Network
Trained by ArtificialImmune System Learning
Fig 3: Actual vs Predicted load Algorithm.12th International Conference on Computer
Modeling and Simulation (UKSim); 2010, p. 408-413.

V. CONCLUSION

In this paper we presented an approach to predict load using

meteorological and holiday values. The existing system
predicts the energy consumption only based on historical load
consumption data. It does not consider the meteorological
factors. Studies have shown that meteorological data also
plays a major role in the load consumption. Also existing
system fails to determine the load which may fluctuate since
they depend only on the known pattern. In this paper we
propose a model which is capable of learning patterns and
making prediction without degrading the accuracy. Results
obtained assert that the proposed model have the potential for
load forecasting which make it suitable for real time
applications.

VI. ACKNOWLEDGMENT

The authors would like to thank everyone for making this a

success. We would like to thank Ms. Amrutha Sasi, Asst.
Engineer (GRID), Load Dispatch Center, Kalamasery, Kerala
for her guidance and help to get the required electricity
consumption data.

REFERENCES

[1]. Navjot Kaur and Amrit Kaur.Predictive Modelling

approach to Data Mining for forecasting electricity
consumption;2016 IEEE.
[2]. Jiawei Han, Micheline Kamber.Data Mining: Concepts
and Techniques. Morgan Kaufmann publications. 2nd
ed.;2006

IJISRT17SP221 www.ijisrt.com 301